Image Categorization Using Local Probabilistic Descriptors

Image categorization involves the well known difficulties with different visual appearances of a single object, but introduces also the problem of within-category variation. This within-category variation makes highly distinctive local descriptors less appropriate for categorization. In this paper we propose a family of local image descriptors, called probabilistic patch descriptors (PPDs). PPDs encode the appearance of image fragments as well as their variability within a category. PPDs extend the usual local descriptors by modelling also the variance of the descriptors' elements, e.g. pixels or bins in a histogram. We apply PPDs to image categorization by using machine learning where the features are the matching scores between images and PPDs. We experiment with two variants of PPDs that are based on complementary local descriptors. An interesting observation is that combining the two PPD variants improves categorization accuracy. Experiments indicate benefits of modelling the within-category variation and show good robustness with respect to noise